Modeling Survival Data with a Cure Fraction

17 concepts13 papers2 proposals

Standard survival analysis is built on a somber mathematical certainty: if you wait long enough, the event of interest will eventually happen to everyone. In technical terms, the survival function $S(t)$—the probability that an individual has not yet experienced the event—is assumed to approach zero as time $t$ goes to infinity.

However, in many real-world scenarios, this "eventual event assumption" is fundamentally flawed. In oncology, a subset of patients may be effectively cured and will never experience a relapse. In digital marketing, as noted by Sinha et al., many users will simply never open a specific email, no matter how long the observation window lasts. When a significant portion of the population is "immune" or "non-susceptible," traditional models produce biased estimates and fail to capture the underlying dynamics of the data.

The Intuition: Disentangling Whether from When

The core innovation of cure fraction modeling is the explicit separation of two distinct processes: susceptibility and timing. Instead of asking only "when will the event happen?", these models first ask "will the event happen at all?"

As Nieto‐Barajas and Engelhard et al. describe, this is typically achieved through a mixture model. The population is treated as a combination of two latent groups:

  1. The Cured Fraction: A group with a zero probability of experiencing the event ($S(t) = 1$ for all $t$).
  2. The Susceptible Fraction: A group that follows a standard survival distribution, where the event will eventually occur.

By decoupling "whether" from "when," researchers can identify which factors influence the likelihood of being cured versus which factors accelerate or delay the event for those who are susceptible. For example, Wu and Yin utilize a logistic regression component to estimate the probability of belonging to the cured group, while simultaneously modeling the survival distribution of the remainder.

Why Standard Models Fail

When a cure fraction exists, traditional survival models like the Cox Proportional Hazards model struggle because they treat long-term survivors as "right-censored" data—individuals who just haven't experienced the event yet.

Wang and Liang argue that this misclassification leads to a loss of valuable information and biased risk estimates. If a model assumes everyone is susceptible, it will underestimate the risk for the truly susceptible group by averaging them with the immune population. Furthermore, Raykar et al. and De Kock and Vlachos point out that time-to-event data is often highly skewed; forcing these distributions into standard frameworks that ignore the "cured" spike at infinity results in poor predictive performance.

Flexible and Neural Approaches

Early work in this field focused on establishing mathematical rigor. Zeng and Ibrahim introduced semiparametric transformation models that provided a flexible, biologically motivated way to handle cure fractions while maintaining optimal efficiency. Similarly, Yin extended these concepts to "frailty models," which account for correlations within subjects, such as multiple teeth in a single patient.

More recently, the field has moved toward integrating machine learning to handle high-dimensional data and complex non-linear relationships:

  • Neural Networks: Engelhard et al. proposed Neural Conditional Event Time (CET) models, which use a binary stochastic layer to explicitly gate the event occurrence. This allows the model to learn complex representations of susceptibility that traditional linear models might miss.
  • High-Dimensional Selection: In genomics, where researchers must sift through thousands of biomarkers, Kızılaslan and Swanson developed penalized mixture cure models to perform variable selection, identifying which specific genes predict a "cure" versus a delayed onset of disease.

Applications Beyond Medicine

While the terminology of "cures" and "susceptibility" originates in clinical trials—such as the COVID-19 studies analyzed by Zhang et al.—the logic applies to any domain with non-responsive sub-populations. Whether predicting if a user will churn from a service, if a conversation will turn into a personal attack (De Kock and Vlachos), or if a customer will ever open a marketing email (Sinha et al.), cure fraction modeling provides a principled way to account for the diversity of human behavior. By acknowledging that some events will simply never happen, these models offer a more accurate and nuanced reflection of reality.

Key Foundational Work

Bayesian Cure Rate Frailty Models with Application to a Root Canal Therapy Study (2005) — contributed 2 concept elements to this direction.
Semiparametric Transformation Models for Survival Data With a Cure Fraction (2006) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_cure_fraction_modeling_77c326e8,
  title = {cure fraction modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cure-fraction-modeling/rdr_77c326e8ebfe4ff08880d1799b7e009c},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_cure_fraction_modeling_77c326e8_v1,
  title = {cure fraction modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cure-fraction-modeling/rdr_77c326e8ebfe4ff08880d1799b7e009c/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake